Protostar Challenge 2026 · Climate Risk Intelligence

Will Everyone Endure the Heat?

Many U.S. cities face heat they were not built for, and the families least able to pay for cooling feel it first.

Team Unicorns · code and data on GitHub

Largest cities in need

    Hover a dot for the city, its homes without AC, its hot days and its low-income bill.

    Definitions

    Two lines decide whether a home can endure the heat: what a family can afford, and what a body can bear.

    Affordable
    6% of income on home energy · median state, single person, bill ÷ bar
    Livable
    82 °F indoors at most · hot nights a year
    Cooling required
    95 °F outside: fans stop protecting · days a year
    “ A home is affordable when a household at the poverty line spends no more than 6% of its income keeping the power on, and livable when it stays at or below 82 °F inside. ”

    APPRISE / ACEEE energy-burden line · U.S. rental cooling codes · CDC and WHO fan limits

    By these lines, a low-income single person was already over the affordable share in 45 of 46 states before any warming.

    Historical Data: Temperature

    U.S. summers already run hotter than the grid was built for, and nights are warming fastest.

    Trends

    • Cooling degree days: +19% in the median city, 1976–2005 to the 2019–2025 forecast; up in 193 of 197 cities.
    • Days with a high of 95 °F or more: +21%. Nights with a low of 75 °F or more: +48%.

    Findings

    • The rise lands where it is already hot: Texas, the Gulf Coast, the Southwest.
    • Our forecast ran low against what happened, so these numbers are conservative.

    Historical Data: Energy Use

    Every 1 °F hotter day adds about 2% to peak electricity demand, in every U.S. region.

    Trends

    • U.S.: +13,700 MW of daily peak per 1 °F of the daily high (2.2%), summers 2015–2018.
    • Most sensitive: Florida, 3.4% per °F. Least: the Northwest, 1.2%. Warmer nights lift the overnight floor about 1.5% per °F.

    Findings

    • Texas's longer record (2002–2018) agrees: 1.9% per °F.
    • Four regions also peaked in winter in 2016–2018: grid risk runs both ways.

    Historical Data: Energy Affordability

    Before any warming, low-income electricity bills were already over the affordable line in 45 of 46 states.

    Trends

    • Real price rose from 11.8¢ (2002) to 12.8¢ per kWh (2018); the average home used about 11,000 kWh a year.
    • Hotter years did not raise prices: −0.002¢ per 100 extra cooling degree days.

    Findings

    • Single person, low income: over the bar in 45 of 46 states. Family of four: 36 of 46.
    • Worst where prices are high (New England, New York, California) or use is high (the South).

    What the Three Datasets Tell Us

    Hot summers raise how much power people use; what it costs follows fuel markets, not the weather.

    Trends

    • Year to year, a hotter summer means more use per home (r = 0.60 for year-to-year changes, 2002–2018).
    • Price does not rise in hot years (r = −0.36); use per home drifted down after 2010 as appliances got more efficient.

    Findings

    • Unaffordability is structural: income, price and baseline use.
    • Heat's cost lands on the grid at the peak, where reserves are planned on past weather.

    Projected Numbers, 2019–2025

    Summers keep getting hotter, but efficiency outpaces the extra cooling, and nothing in the record says prices will follow the heat.

    Trends

    • Weather: cooling degree days average 1,531 in 2019–2025, +8% on 2002–2018 (1,420).
    • Use per home: the extra heat adds about 95 kWh a year (+0.9%), but the efficiency trend (−28 kWh a year) brings use down to about 10,700 kWh.

    Findings

    • Price is held at 2018's 12.8¢: in 2002–2018 it moved with fuel markets, not weather.
    • So the average bill barely moves; the heat shows up at the peak and in the regions, not in the national average.

    Regional Affordability, 2019–2025

    Heat strain is regional, but unaffordability is everywhere: where it gets hotter is not where bills hurt most.

    Cooling-required days a year

    Low-income family bill ÷ affordable bar

    Heat exposure against the bill, by state

    Findings

    • r: heat exposure barely predicts which states are unaffordable.
    • A low-income family is over the bar in 36 of 46 states, from Maine to Texas.
    • Forecast heat adds only $17–33 a year to the median low-income bill; its bigger cost is on the grid (Texas runs short in a hot summer).

    Our forecast (data up to 2018), EIA-861, RECS 2015/2020, Census LACE, NERC 2018. 2018 dollars.

    The Questions We Are Asking

    Our forecast shows where today’s plans fall short. What to do about it is a choice for the people who hold each lever, so pick a role to see the question we hand them and the evidence behind it.

    The risk we noticed

    More very hot days and hot nights, and hotter afternoon peaks. Compared with 1976–2005, our 2019–2025 forecast has 21% more days of 95 °F or more and 48% more nights of 75 °F or more. The grid feels it at the peak; homes feel it where cooling is missing or costs too much to run.

    Who they are: the groups that decide how much spare power the grid must keep ready for the hottest afternoons, and the utilities that build to meet it.

    What we found

    Grids are planned with a cushion of spare capacity, and that cushion was sized on past weather. By 2025 the warming we forecast uses up about a fifth of it. In a hot summer, Texas’s main grid falls 1.3% short of what people would want to use.

    Demand on the hottest days climbs about 2% for every degree of extra heat, so small shifts in temperature matter more than they sound.

    The questions we are asking

    • Should the spare-capacity target be built on forecast weather instead of the weather of the past?
    • How much extra margin is worth paying for, and where does it matter first?
    • Is it cheaper to build more supply, or to shrink the peak by asking users to use less at that hour?

    Where margins were already thin, the questions are the most urgent.

    The evidence

    • WeatherMore very hot days and hot nights are coming.
    • Energy UseHow much more power each degree of heat costs the grid.
    • Projected NumbersHow much of the cushion warming uses up.
    • LimitationsWhat we could not measure about the grid.
    Next: how we would answer them →

    Our forecast (data up to 2018) and public energy, price and housing data from the federal government · 46 states, 2018 dollars · every number above is detailed in the linked slides

    How We Will Answer It

    One concrete proposal for each group, built so that each helps the others. Pick a role to see what it would do, who else it affects and what supports it.

    Who acts: regional grid operators and the state bodies that approve utility plans.

    Our proposal

    • Plan the spare capacity for the hottest afternoons on forecast weather, not only past weather, by adding a warming allowance to each region’s summer peak estimate.
    • Count peak-cutting as capacity: efficient cooling, and agreements to use less at the peak hour, should compete on equal terms with new power plants.
    • Start where the cushion was already thin, such as Texas.

    What it does for the other sides

    • Households: fewer outages in a heat wave.
    • Rate regulators: the allowance shows up in bills, so cutting the peak is the cheaper first move.
    • Aid funders: efficient cooling in low-income homes becomes a source of capacity, not only a social program.

    The evidence

    • Auffhammer et al. 2017Peak demand is projected to grow faster than average use as it warms.
    • IEA 2018More efficient cooling can cut the need for new power plants at the peak.
    • Energy UseWhat each degree of heat costs the grid.

    The thread that runs through all four

    Use less power at the hottest hour. Efficient cooling in low-income homes cuts the peak the grid must meet, the bill a family must pay, and the number of homes with no cooling at all. Each group can start with powers it already has, and each proposal is built to help the others.

    Next: the drawbacks, and the numbers we still need →

    Proposals are ours; the numbers come from our forecast and public data · papers are cited in full on the last page of this section

    The Drawbacks, and the Numbers We Still Need

    No proposal is free. For each group, here is who else bears a cost, what the research warns, and what we must measure before anyone finalizes a decision.

    Drawbacks and open numbers for this group’s proposal

    Drawbacks, and who else bears them

    • Planning on forecast weather raises costs sooner: customers pay for capacity that is needed only if the warming arrives.
    • If it warms more slowly than forecast, the extra margin sits idle.
    • Peak-cutting counts as capacity only if people really respond when asked.

    Numbers we need to finalize

    • How much extra capacity each region needs by 2025 and 2030, from our forecast.
    • What that costs by building supply versus cutting the peak.
    • How much peak cut efficient cooling and peak-hour agreements really deliver.
    • How uncertain the forecast is, region by region.

    What the research warns

    • Davis et al. 2014Efficient refrigerators saved about a quarter of what was predicted, and replacing air conditioners raised power use in Mexico.
    • Auffhammer et al. 2017Their peak estimates assume today’s technology and prices.

    Before any group decides

    Every group needs the same three numbers: how big the problem is in their own place, what a fix costs per household or per unit of power, and what a small trial actually delivers. We would not scale any piece until the trial results are in.

    Next: why it lasts →

    Papers are cited in full on the last page of this section · “about $300 a year” is the median affordability gap across states, 2019–2025 forecast, 2018 dollars

    Why It Lasts

    A plan is sustainable if it helps now without making emissions or costs worse later. We think this one does, as long as the cooling is efficient.

    Emissions

    Cooling gets more efficient, not just bigger.

    Cooling demand will climb as the world warms and more households install air conditioning (Davis and Gertler 2015). Efficiency is the biggest lever against that: strong efficiency policy could cut the world’s cooling energy use by 45% by 2050 compared with doing nothing (IEA 2018). In 55 U.S. cities, heat pumps reduced emissions damages in about 7 in 10 homes (Deetjen et al. 2021).

    Could fail if the extra power comes from a dirty grid, or old equipment is swapped for a unit that is hardly better.

    Costs

    Cutting the peak is cheaper than building for it.

    Under a high-warming path, peak demand is projected to grow about twice as fast as average use (Auffhammer et al. 2017), and the grid must be sized for the peak. Efficient cooling reduces the need for new power plants at that hour, and the same agency estimates it lowers total investment and running costs by about $3 trillion worldwide to 2050 (IEA 2018).

    Could fail if savings come in below predictions: replaced refrigerators saved a quarter of the predicted amount (Davis et al. 2014), and Michigan home upgrades cost about twice the energy they saved (Fowlie et al. 2018). That is why trials come first.

    Over the long run

    Keeps cooling a safety net as hot days multiply.

    Home air conditioning explains nearly all of the 75% fall in the death toll of very hot days in the U.S. since 1960 (Barreca et al. 2016). With more very hot days and nights coming, that protection holds only if homes have cooling, can afford to run it, and the grid stays on. The four proposals aim at exactly those three conditions.

    Could fail if funding fades, or warming outruns the forecast. We would revisit the plan each year with new forecasts and trial results.

    Figures are the authors’ own: 45%, $3 trillion and the fall in heat deaths are global or national, not city forecasts.

    Limitations

    Warming is a modest part of grid risk to 2025. Here is what we assumed, what our data cannot see, and what would change the answer.

    Where the 2025 grid margin goes, median NERC area

    Percentage points of reserve margin lost

    Warming (our forecast)−1.6
    The plan’s own change (growth, retirements)−2.3
    One hot summer (about 1 in 10)−8.4

    What we assumed

    • NERC’s 50/50 peak forecasts do not already include warming.
    • Demand rises linearly with heat; the peak day’s high stands in for the peak hour.
    • A higher price in scarcity is only a scenario: hot years did not raise prices in 2002–2018.

    What our data cannot see

    • No humidity: dry-heat thresholds understate risk in humid places.
    • Airport stations miss city-centre heat; EIA-930 covers only 3.5 summers.
    • RECS income brackets approximate poverty; Alaska and Hawaii are left out.

    Outside our scope

    • The cold side: four regions hit their yearly peak in winter in 2016–2018. Grid risk runs both ways.
    • Load is what customers drew, so demand cut in emergencies does not show.

    What would change the conclusion

    Plans already include warming. Then the planning gap shrinks.

    The poor cut cooling more. Likely: bills fall but homes get hotter, so livability is worse than shown.

    Our forecast runs low. It was conservative on heat days, so warming’s effects are probably understated.

    docs/policy_essay_outline.md §7 · outputs/grid_projection_national/summary.md · outputs/grid_national/summary.md

    Reproducibility & Openness

    One command rebuilds the submitted forecast byte for byte, and every number in this deck traces back to a script and a committed file.

    1

    The forecast

    • No randomness: same data in, same file out. Package versions pinned in requirements.txt.
    • Rebuilt from a clean copy of the repo: identical to submission.csv (5,880 rows, SHA-256 f1ba20cc…).
    • Training data pinned to revision bbbdebda… in shared.py, so later dataset updates cannot change it.
    pip install -r code/requirements.txt
    python code/exp05b_daily_distribution.py
    python code/check_submission.py submission.csv

    Under a minute on a laptop.

    2

    This deck

    1. Challenge data to 2018 + EIA, NERC, RECS, Census LACE
    2. policy_*.py: each writes its fixed settings to config.json before any result
    3. outputs/*/summary.md and CSVs, all committed
    4. export_deck_data.py → deck_data.js → these slides

    Every intermediate file is in the repo, so any step can be re-run on its own. Sources: data/README.md, data/policy_sources.md.

    3

    The data rule, in code

    • assert_cutoff stops any run that sees a year after 2018.
    • Each experiment's decision rule is written to decision.json before its forecast.
    • 2019–2025 observations: used only to score frozen forecasts after the fact, declared in the README, and kept out of these slides.
    • Agent rules in code/AGENTS.md; every AI step is listed in the README.

    Known gaps, stated up front

    Raw downloads. Agency files for the policy analysis are cached in data/raw/, which is not committed; a full rebuild re-downloads them.

    Census key. County matching needs a free Census API key. Its output is committed, so later steps run without one.

    Old scores. Held-back cdd scores in Experiments 3–7 predate the valid_flags fix. Forecasts are unaffected.

    Code, data and outputs: github.com/JessCai06/-Protostar-Challenge-CMU-2026-Climate-Risk-Intelligence · README “Reproduce”

    Conclusion

    Not everyone will endure the heat on plans built for yesterday’s weather.

    Grids and household budgets are planned on past summers. Our forecast for 2019–2025, made only from data up to 2018, shows where that falls short.

    +48%

    The heat is uneven

    Hot nights rise by half and cooling-required days by a fifth, mostly in Texas, the Gulf Coast and the Southwest. Nights warm faster than days.

    ⅕

    The grid feels it first

    Warming uses up about a fifth of the planned reserve cushion. ERCOT, already below its reference margin, runs short in a hot summer.

    45 of 46

    Bills were already too high

    Low-income electricity bills were over the affordable bar in 45 of 46 states before any warming. Heat adds only $17–33 a year.

    So

    Plan reserves on forecast weather, fix affordability where it is structural, and bring cooling to homes that have none. One lever serves all three: efficient cooling for low-income homes, which lowers the peak, lowers the bill and makes homes livable.

    Team Unicorns · Protostar Challenge 2026 · Code, data and the full essay: github.com/JessCai06/-Protostar-Challenge-CMU-2026-Climate-Risk-Intelligence